The Self, Cognizant of Its Relations with Others in Time
ASADA--With robots, if the inherent traits and coding are clearly written they will behave within the defined parameters, but if you even vary a bit they will become immobilized. This is one definition of a robot that functions as it is supposed to, but without a bit more plasticity, individuality and the capacity to feint in play, etc. will not emerge. I really would like my robots to evolve to this level.
SAKURA--I know what you mean. Ethologist Konrad LORENTZ said it originally, but I, too, consider the evolution of life to be one process of learning. Genetics adapting to the environment itself can be considered one "learning." But the learning in one individual's lifetime--by a typical definition of "learning"--is impossible without some mechanism for storing the information in the central nervous system. Even cases of making robots learn will become increasingly difficult, don't you think?
ASADA--We are using mostly reinforcement learning,[*11] but without a concept of time, it doesn't turn into memory. So my problem is really how to make them gain a concept of time. They can understand a given number of seconds, but not longer sequences, or more to the point an apparent application of sequential structure. They're still completely time reactive.
SAKURA--And this is what's difficult?
ASADA--It all depends on how you approach it. I would like to work out an effective structure of experiences and learning processes that would enable the robots to understand how concepts of time are arrived at. If they are presented with a space they can apprehend three-dimensional information and, therefore, project themselves into it and understand their spatial relation to things within that space. But time is more difficult. I've physically put time pieces inside of them, but they still have no subjective notion of time. They're only dealing with increments. And I don't want them to deal with increments, but learn to evolve an understanding of their own position in relation to that continuum.
SAKURA--So your dilemma is to resolve, for shooting robots, how to learn a sense of time, subjective time, while at the same time living with a time mechanism planted inside of them.
ASADA--They have a beating heart, which structures their activities, but I haven't specified concepts like yesterday, or the day before. I'd like them to understand that yesterday they came so far, and then today even further, to give structure and form to the sequence of their activities. Then they will be able to speak about their own pasts.
SAKURA--But doesn't this exceed the scope of RoboCup? It is only my intuition, but even within the history of evolution this is something that never emerged to any precise degree. It is a product of human culture, institutions like education. It was only after writing and then mathematics emerged that concepts like "yesterday" and "today" developed. I don't believe that it is something that evolved naturally.
ASADA--Yes, I've been told that by many people that our sense of time is quite specifically human.
SAKURA--Talking about Africa once again, I used to purchase fruits from a local vendor, and I often was short on small change, so I'd have a tab running. Then one day, the fruitseller told me that I hadn't paid for five bananas from "yesterday." Looking at my records, I wasn't down for five bananas the day before or even the day before that. We got into a bit of a heated discussion about it, but it came out that I had put five bananas on credit some two weeks earlier and forgotten to pay. It was there in his ledger. In their vocabulary, time was either yesterday, today or tomorrow. (laughs) There was absolutely no distinction between one and two days previous.
Of course, it was like this everywhere else on earth until quite recently. In Japan, in the Heian Period, if a man wanted to marry a woman, he would court her for three days. It was considered a marriage process, and they would hold a "Third Day Rice Cake Ceremony." This "third day" was just a symbolic way of expressing a time greater than the present, an abstraction for an eternity, a continuum. All over the world, the emergence of precise concepts of time came only after capitalism. And if that is the case, then trying to get your robots to apprehend their selves of today versus their selves of yesterday . . . .
ASADA--The question of how to apprehend time is something that I'd like them to think out for themselves. It all depends on how a robot would approach the concept of time.
SAKURA--And when you reach this point, will distinctions like the distant past, recent past, the present day, near future, and distant future be enough? Or will you attempt to solve issues like 10 years ago, 3 months ago, 1 week ago, day before yesterday, yesterday, today, tomorrow, day after tomorrow . . . . Will such distinctions be necessary?
ASADA--This may be an extreme way to describe it, but they are going to need a good enough structural grasp of the concept to be able to describe how they grew up, if they are to develop a sense of identity.
SAKURA--Who knows which will come first? Their sense of identity being unchanging and . . . .
ASADA--That is the point. The fact that they see themselves as having been there throughout their existence, and understanding the same about the other. In order to be aware of this they have to be able to codify time.
SAKURA--I would imagine that "yesterday" and "today" are concepts that will need to be taught, "top-down"[*12] learning if you will, though less specific. "Previously, presently, and from now" seem achievable aims. Chimpanzees and gorillas have this much grasp of time. But as I said before, the problem is having the tools for more refined concepts.
ASADA--At least let us agree that the concept of self is inextricably linked with some conception of time. The other is always the other, and that this "other" constant is relevant to understanding one's self as constant is the same issue in a very essential way.